3 ms·

Yes, ChatGPT seems to strongly prefer happy endings. Ask it to write a scene where a Montague bumps into a Capulet, and accusations start flying. One of the cha

by ekidd 4y ago

Yes, ChatGPT seems to strongly prefer happy endings. Ask it to write a scene where a Montague bumps into a Capulet, and accusations start flying. One of the characters will successfully de-escalate. Or the price of Verona will show up with guards and stop the brawl. He'll even throw in some foreshadowing about how this will all end in tragedy. I suspect that this relentless love of upbeat endings is a side effect of the RLHF training used to create the "assistant" persona. They wanted a friendly and harmless chatbot. And apparently it "likes" exactly the kind of stories you'd expect.

professoretc 4y ago
I've noticed a similar phenomenon when asking it to solve problems; if you point out a "flaw" in its proposed solution, even something that isn't wrong at all, it will cheerfully agree with you and "fix" its proposed solution, often making it worse or wrong in the process.
- tharkun__ 4y ago
  Good news everyone! (Futurama reference) Our jobs are safe after all! The junior dev helped by ChatGPT will not take your job. ChatGPT cheerfully agreed with the junior when he didn't understand something and thought to point out how this can't be right and ChatGPT gave him a "fix" that brought down Production.
lostlogin 4y ago
> ChatGPT seems to strongly prefer happy endings I always thought this a Hollywood thing. “Wrap up the loose ends and make it a happy ending.” Movies and series that end untidily are often my favourites.
BizarroLand 4y ago
I was using koboldAI to work on a short story and it had two characters bump into each other, one of them pull a knife, viciously shank the other seven times leaving the other one to collapse into a pool of blood and die, and then apologize for the accident like it was a minor occurrence after which they both went on their own ways with a "slightly embarrassed smile and nod, wistfully wondering of what might have been".

by ekidd 4y ago

professoretc 4y ago

I've noticed a similar phenomenon when asking it to solve problems; if you point out a "flaw" in its proposed solution, even something that isn't wrong at all, it will cheerfully agree with you and "fix" its proposed solution, often making it worse or wrong in the process.

tharkun__ 4y ago
Good news everyone! (Futurama reference) Our jobs are safe after all! The junior dev helped by ChatGPT will not take your job. ChatGPT cheerfully agreed with the junior when he didn't understand something and thought to point out how this can't be right and ChatGPT gave him a "fix" that brought down Production.

lostlogin 4y ago

> ChatGPT seems to strongly prefer happy endings I always thought this a Hollywood thing. “Wrap up the loose ends and make it a happy ending.” Movies and series that end untidily are often my favourites.

BizarroLand 4y ago

I was using koboldAI to work on a short story and it had two characters bump into each other, one of them pull a knife, viciously shank the other seven times leaving the other one to collapse into a pool of blood and die, and then apologize for the accident like it was a minor occurrence after which they both went on their own ways with a "slightly embarrassed smile and nod, wistfully wondering of what might have been".